Dr. Yiqun Xie Received the Best Paper Award at the 2021 IEEE International Conference on Data Mining

Assistant Professor Dr. Yiqun Xie received the Best Paper Award with collaborators for their work: "A Statistically-Guided Deep Network Transformation and Moderation Framework for Data with Spatial Heterogeneity," at the 2021 IEEE International Conference on Data Mining (ICDM).

Background: IEEE ICDM is a top data mining venue in the computing field, and has a highly-selective peer-review process (e.g., 9.9% acceptance rate for regular full-length papers). The conference received 990 full-length submissions in 2021. Xie's work is also invited to be presented at the Sister Conference Best Paper Track of IJCAI-ECAI 2022, a premier artificial intelligence forum. Conferences in general computing research are considered as important as journals based on both selectivity and impact, as illustrated in this National Academy Press article and this page featuring an informative video on the detailed peer-review process.

Xie's research tackles a fundamental challenge posed by spatial data to general machine learning approaches: spatial heterogeneity (e.g., see a recent PNAS article). As the processes/functions from observed features (e.g., spectral values from satellite imagery) to prediction targets (e.g., land cover classes) are often nonstationary over space — which violates the common identical distribution assumption underlying machine learning methods — attempts that train a model with a single set of parameters for different regions may lead to "hit or miss" solutions and significantly limit model performance. Making the problem more challenging, the spatial footprints of different processes are often unknown in practice.

The paper offers a new approach that addresses this key challenge by statistically-guided spatial transformation, which can automatically recognize heterogeneous spatial processes in the input data and capture their footprints. It also introduces a spatial moderator that can generalize heterogeneity patterns learned in one region to a new location. The framework can be applied to general deep learning architectures, and experiment results have demonstrated substantial performance improvement on fully-connected (e.g., ANN) and recurrent (e.g., LSTM) neural networks.

Published on Wed, 02/02/2022 - 18:09